DSO: Dual-Scale Neural Operators for Stable Long-term Fluid Dynamics Forecasting
In the realm of scientific research and engineering, long-term fluid dynamics forecasting is a pivotal challenge. The ability to accurately predict fluid behavior over extended periods is essential for numerous applications, ranging from weather forecasting to aerospace engineering. Recent advancements in artificial intelligence, particularly through the use of neural operators, have shown promise in modeling systems governed by partial differential equations (PDEs). However, these models often encounter significant hurdles in terms of long-term stability and precision.
A recent study, documented in the preprint arXiv:2603.26800v1, introduces a novel approach to address these challenges through the development of the Dual-Scale Neural Operator (DSO). This innovative model confronts two primary failure modes observed in existing architectures:
- Local Detail Blurring: Fine-scale structures, such as vortex cores and sharp gradients, are progressively smoothed out, resulting in a loss of critical information necessary for accurate predictions.
- Global Trend Deviation: The overall motion trajectory of the fluid tends to drift away from the actual ground truth during prolonged forecasting periods, leading to significant inaccuracies.
The authors of the study assert that these failures stem from the uniform treatment of local and global information processing in current neural operators. In reality, local and global characteristics evolve differently within physical systems, necessitating a more nuanced approach. To bridge this gap, the DSO model integrates two complementary modules:
- Depthwise Separable Convolutions: This component focuses on fine-grained local feature extraction, allowing the model to capture intricate details in the fluid dynamics effectively.
- MLP-Mixer: This module is designed for long-range global aggregation, enabling the model to maintain awareness of broader motion trends that influence the overall dynamics.
The effectiveness of the DSO model is supported by extensive numerical experiments conducted on vortex dynamics. These experiments reveal that nearby perturbations predominantly impact local vortex structures, while distant perturbations have a significant effect on global motion trends. This empirical validation underscores the rationale behind the dual-scale approach adopted by DSO.
Furthermore, the DSO model has been rigorously tested against established benchmarks in turbulent flow forecasting. The results demonstrate that DSO achieves state-of-the-art accuracy while ensuring robust long-term stability. Notably, it reduces prediction error by over 88% when compared to existing neural operators, marking a substantial improvement in the field of fluid dynamics forecasting.
As the demand for accurate long-term forecasting in fluid dynamics grows, innovations like the Dual-Scale Neural Operator represent a significant step forward. The DSO framework not only enhances predictive capabilities but also opens avenues for further research and application in various scientific and engineering domains.